Stable Diffusion 3.5 beginner tutorial: install, model selection by VRAM (6GB minimum), first image in AUTOMATIC1111 or ComfyUI, prompt tips, ControlNet basics, and a common-error table..

💡 What You Will Learn

Stable Diffusion 3.5 beginner tutorial: install, model selection by VRAM (6GB minimum), first image in AUTOMATIC1111 or ComfyUI, prompt tips, ControlNet basics, and a common-error table..

📜 Table of Contents

• 1. Preparation: Hardware Requirements and Two Tools
• 2. Installation: Two Approaches
• 3. Model Selection: Large, Medium, or Turbo
• 4. Generating Your First Image: 5 Steps
• 5. Advanced Control: The ControlNet Four-Piece Set
• 6. Quick Reference for Common Errors
• FAQ
• Related Articles

2026 Stable Diffusion 3.5 Beginner's Tutorial: From Installation to Generating Images

AI image generation tools are getting more powerful, but many people are still stuck at the "using someone else's online service" stage — they want to run it locally but worry about the hassle. Stable Diffusion 3.5 has surpassed 48 million downloads on HuggingFace, making it the most popular open-source image generation model, and it can run with as little as 6GB of VRAM. This tutorial follows the order of "set up environment → choose model → generate first image → advanced control" to get you fully up and running with SD3.5.

1. Preparation: Hardware Requirements and Two Tools

Running SD3.5 locally doesn't have a high hardware bar:

• Minimum specs: 6GB VRAM GPU — it runs, but slowly and with model limitations
• Recommended specs: 8-16GB VRAM — SD3.5 Medium and most features run smoothly
• Not enough VRAM?: You can add launch parameters to reduce VRAM usage (e.g., low VRAM mode), trading speed for capacity

Which tool should you pick? The two mainstream open-source interfaces right now are:

Beginners are advised to start with AUTOMATIC1111: the interface is intuitive, answers are easy to find, and any error you run into can be solved with a quick search.

2. Installation: Two Approaches

Option A: All-in-one package (recommended for beginners)

Community all-in-one packages work right after extraction — they come with models, dependencies, and a launcher built in, so you don't need to install Python or CUDA. The downside is that version updates lag behind, and troubleshooting can be trickier when something goes wrong. It's best for getting a feel for things first.

Option B: Official manual installation

# Requires Python 3.10+ and Git
git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui
cd stable-diffusion-webui
# On Windows, run webui-user.bat — it auto-installs dependencies and launches

After startup, open http://127.0.0.1:7860 in your browser to access the interface. The first launch will download dependencies, so be patient for a few minutes.

3. Model Selection: Large, Medium, or Turbo

SD3.5 comes in several versions — you need to match VRAM with results:

After downloading a model, place it in the models/Stable-diffusion/ directory (the models folder in WebUI), then select it from the model dropdown at the top of the interface.

4. Generating Your First Image: 5 Steps

1. Select the SD3.5 model from the dropdown at the top
2. Write a clear positive prompt describing the image content: a cute orange cat astronaut on mars, cinematic lighting, 8k
3. Write a negative prompt for things you don't want: blurry, low quality, deformed hands
4. Keep sampling steps at the default 20-30, and set the dimensions as needed (e.g., 1024×1024)
5. Click "Generate" and wait for the progress bar to finish — the image appears on the right

Prompt formula: subject + scene + style + lighting + quality keywords, separated by English commas. Chinese prompts work too, but English generally produces more consistent results.

5. Advanced Control: The ControlNet Four-Piece Set

To get the AI to generate images strictly according to your vision, you need ControlNet. It uses additional inputs (line art, poses, depth maps) to constrain the generation process:

• Canny (edge detection): Give it a line drawing, and the AI colors within the contours — great for redrawing and coloring
• OpenPose (pose recognition): Provide a human pose skeleton, and the AI generates a character in the same pose — great for character design
• Depth (depth map): Preserves the spatial hierarchy of a scene, letting the AI swap content within the original composition — great for scene reconstruction
• IP-Adapter (style transfer): References the style of one image so new images maintain the same look

In WebUI, ControlNet is a plugin. Once installed, each feature corresponds to a separate model, and the workflow is always "upload reference image → select feature → adjust weight → generate."

6. Quick Reference for Common Errors

• Out of VRAM (CUDA out of memory): Switch to the Medium model, or add low VRAM mode to your launch parameters
• Black or garbled images: VAE settings are wrong or the model file is corrupted — check model integrity
• Generation is too slow: Lower resolution, reduce steps, or close background programs that hog VRAM

FAQ

Q: Can I run it with 6GB of VRAM? A: Yes. Using the Medium model with low VRAM mode will work, though generation will be slower. When VRAM is tight, prioritize getting images to generate first, then optimize speed gradually.

Q: Can I use generated images commercially? A: The open-source version of SD3.5 follows its open-source license. Before commercial use, confirm the license terms of the specific version you downloaded — restrictions vary by version, so defer to the official documentation.

Q: Do I need to install both WebUI and ComfyUI? A: No. Both tools share the same model files. Get comfortable with WebUI first, then migrate to ComfyUI when you need fine-grained control (batch processing, workflow reuse) — they don't conflict.

Q: Why do my generated images look so different from others'? A: It's most likely a prompt or model issue. Start by reproducing the official example prompts to confirm your environment is fine, then gradually add your own descriptions. Also make sure your model version matches — don't apply SD1.5 prompt styles to SD3.5.

Related Articles

• ControlNet Guide 2026: Pose, Depth, Canny — A Detailed Look at Stable Diffusion Control Methods
• AI Image Generation 2026: Getting Started with Stable Diffusion and ComfyUI
• 2026 Stable Diffusion Beginner's Tutorial